Targets and Industries

Kerry Back

Transforming the target variable

We should usually deal with outliers for the target variable.
We’ve used ranks, but there are more systematic ways.
For example, we could use QuantileTransformer.
When we predict with the model, we get “untransformed” predictions (predicting original target)

Usual example

roeq and mom12m in 2021-12
RandomForestRegressor
transform1, poly, transform2
Use GridSearchCV to find max_depth

Changes

Instead of model=RandomForestRegressor, use model = TransformedTargetRegressor(…)
- input1: regressor
- input2: transformer
param_grid = {“transformedtargetregressor__regressor__max_depth”: [4, 6, 8]}
pipe = … (as before)
cv = … (as before)

from sklearn.compose import TransformedTargetRegressor

transform3 = QuantileTransformer(
    output_distribution="normal"
)

model = TransformedTargetRegressor(
    regressor=RandomForestRegressor(random_state=0),
    transformer=transform3
)

pipe = make_pipeline(
    transform1,
    poly,
    transform2,
    model
)

param_grid = {
    "transformedtargetregressor__regressor__max_depth": [4, 6, 8]
}

cv = GridSearchCV(
    pipe, 
    param_grid=param_grid
)

X = data[["roeq", "mom12m"]]
y = data["ret"]

cv.fit(X, y)

Exercise

Run GridSearchCV with

MLP Regressor,
hidden layer sizes = (16, 8, 4, 2) and (8, 4, 2) ,
and a transformed target variable.

Industries

Used OneHotEncoder to create dummy variables
Let’s also use deviations from industry means as predictors
E.g., maybe we want to buy stocks with high ROEs relative to their industry rather than just high ROE stocks in general

Calculating deviations from means

One industry at a time:

data["roeqx"] = data.groupby("industry").roeq.transform(
    lambda x: x - x.mean()
)
data["mom12mx"] = data.groupby("industry").mom12m.transform(
    lambda x: x - x.mean()
)

Or in a loop:

for char in ["roeq", "mom12m"]:
    data[char+"x"] = data.groupby("industry")[char].transform(
        lambda x: x - x.mean()
    )

Industry momentum

We could add industry means as additional features.
Probably most important for momentum. Industry momentum seems to be a significant phenomenom.
Could compute momentum of industry returns. But average momentum within the industry is easier and probably close.

data["mom12mi"] = data.groupby("industry").mom12m.mean()

Two versions: with and without industry dummies

In both versions:

transform3 = QuantileTransformer(
    output_distribution="normal"
)
model = TransformedTargetRegressor(
    regressor=RandomForestRegressor(random_state=0),
    transformer=transform3
)

Without dummies:

pipe = make_pipeline(
    transform1,
    poly,
    transform2,
    model
)
X = data[["roeq", "mom12m", "roeqx", "mom12mx", "mom12mi"]]

With dummies

numeric_transform = make_pipeline(
  transform1,
  poly,
  transform2
)

category_transform = OneHotEncoder()

feature_transform = make_column_transformer(
  (numeric_transform, ["roeq", "mom12m", "roeqx", "mom12mx", "mom12mi"]),
  (category_transform, ["industry"])
)

pipe = make_pipeline(
  feature_transform,
  model
)

X = data[["roeq", "mom12m", "roeqx", "mom12mx", "mom12mi", "industry"]]